Input-Shape Fixed Image Compression for Faster Neural Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Learning-based image compression using neural networks is computationally expensive, leading to increased resource usage costs and processing time, which offsets the cost savings from reduced data storage and transfer, and optimizing neural nets for each image size requires time and resources.
Innovation Solution
A data processing method that generates input shape fixed compressors and decompressors to process images of varying sizes efficiently, using a combination of general-purpose and input/output shape fixed neural networks, optimizing hardware resources for faster and cheaper processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If learning-based compression processing is performed using a neural network, then image quality is improved, but processing time increases
Solution Approach 1:
The patent pre-generates multiple input shape fixed compressors for different image sizes before actual compression processing. This preliminary preparation allows the system to directly apply the appropriate pre-trained compressor without time-consuming optimization during runtime, thus reducing processing time while maintaining high image quality through neural network-based compression.
Solution Approach 2:
The system dynamically selects the most appropriate input shape fixed compressor based on the actual size of the input image. By having multiple compressors optimized for different shapes and dynamically choosing the best match, the system achieves both high compression quality and efficient processing without requiring real-time neural network optimization.
2Loss of time
If learning-based compression processing is performed using a GPU, then processing time is reduced, but resource usage cost increases
Solution Approach 1:
The system performs the computationally expensive neural network optimization work in advance during an offline training phase, generating multiple input shape fixed compressors for different image sizes. During actual compression processing, the system only needs to apply these pre-optimized compressors, which significantly reduces runtime computational requirements and resource usage costs while maintaining fast processing.
3Loss of time
If a library optimizing inference processing is used, then processing time is reduced, but input size must be fixed
Solution Approach 1:
The patent segments the compression task into multiple specialized compressors, each optimized for a specific input shape. Instead of using a single general-purpose compressor that handles all sizes inefficiently, the system divides the work into specialized components (input shape fixed compressors for different sizes) that can be efficiently selected and applied based on the actual input image dimensions.
Solution Approach 2:
The system creates a universal compression solution by generating multiple input shape fixed compressors that collectively handle various image sizes. This multi-functional approach allows the system to maintain the efficiency benefits of fixed-size optimization while preserving the ability to process images of different dimensions by selecting the appropriate specialized compressor.
4Measurement precision
If a neural net optimized for each image size is generated, then compression quality is improved, but optimization time increases
Solution Approach 1:
The system performs the time-consuming neural network optimization work in advance during an offline training phase, generating multiple input shape fixed compressors for different image sizes. By moving the optimization work to a preliminary stage, the system achieves high compression quality through size-specific optimization without incurring optimization time delays during actual compression processing.
Data Source
AI summary
In a data processing method executed by a data processing system that performs compression and/or decompression of image data, a tensor shape representing compression target data is obtained, compression processing is performed using data having an input shape fixed for each shape of the compression target data as input, and an input shape fixed compressor for outputting compressed data is generated. Then, the data processing system performs compression processing of compression target data using the generated input shape fixed compressor to generate compressed data.